Selection of Variables in Small Business Failure Analysis

This paper focuses on one of the most determinant processes in business failure assessment: variable selection. We apply first-level variable selection based on previous literature on SMEs. Next, we perform a statistical variable selection on a sample of 3210 small firms using both mean and median differences. As the resulting variables differ, we run a varied group of business failure assessment methods (linear discriminant analysis, quadratic discriminant analysis, logistic discriminant analysis, kth-nearest-neighbor discriminant analysis, logit, probit, and data envelopment analysis) to identify the implications of using both tests of differences. Our results show that median differences outperform mean differences in selecting variables, and non-parametric methods outperform parametric ones in identifying failed and not-failed firms by avoiding the effects of dispersion and skewness. Additionally, we contribute new evidence on the addition of qualitative information (payment incidents), with previous evidence for SMEs being scarce